gru: support for asynchronous gru instructions
Jack Steiner [Wed, 17 Jun 2009 23:28:25 +0000 (16:28 -0700)]
Add support for asynchronous GRU instructions.  Currently, asynchronous
instructions are supported only for GRU instructions issued by the kernel.

[akpm@linux-foundation.org: build fix]
Signed-off-by: Jack Steiner <steiner@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

drivers/misc/sgi-gru/grufault.c
drivers/misc/sgi-gru/grukservices.c
drivers/misc/sgi-gru/grukservices.h

index 3220e95..8443e90 100644 (file)
@@ -468,10 +468,6 @@ irqreturn_t gru_intr(int irq, void *dev_id)
                return IRQ_NONE;
        }
        get_clear_fault_map(gru, &imap, &dmap);
-       gru_dbg(grudev,
-               "irq %d, gid %d, imap %016lx %016lx, dmap %016lx %016lx\n",
-               irq, gru->gs_gid, dmap.fault_bits[0], dmap.fault_bits[1],
-               dmap.fault_bits[0], dmap.fault_bits[1]);
 
        for_each_cbr_in_tfm(cbrnum, dmap.fault_bits) {
                complete(gru->gs_blade->bs_async_wq);
index a0f9810..9dff33c 100644 (file)
  * loaded on demand & can be stolen by a user if the user demand exceeds the
  * kernel demand. The kernel can always reload the kernel context but
  * a SLEEP may be required!!!.
+ *
+ * Async Overview:
+ *
+ *     Each blade has one "kernel context" that owns GRU kernel resources
+ *     located on the blade. Kernel drivers use GRU resources in this context
+ *     for sending messages, zeroing memory, etc.
+ *
+ *     The kernel context is dynamically loaded on demand. If it is not in
+ *     use by the kernel, the kernel context can be unloaded & given to a user.
+ *     The kernel context will be reloaded when needed. This may require that
+ *     a context be stolen from a user.
+ *             NOTE: frequent unloading/reloading of the kernel context is
+ *             expensive. We are depending on batch schedulers, cpusets, sane
+ *             drivers or some other mechanism to prevent the need for frequent
+ *             stealing/reloading.
+ *
+ *     The kernel context consists of two parts:
+ *             - 1 CB & a few DSRs that are reserved for each cpu on the blade.
+ *               Each cpu has it's own private resources & does not share them
+ *               with other cpus. These resources are used serially, ie,
+ *               locked, used & unlocked  on each call to a function in
+ *               grukservices.
+ *                     (Now that we have dynamic loading of kernel contexts, I
+ *                      may rethink this & allow sharing between cpus....)
+ *
+ *             - Additional resources can be reserved long term & used directly
+ *               by UV drivers located in the kernel. Drivers using these GRU
+ *               resources can use asynchronous GRU instructions that send
+ *               interrupts on completion.
+ *                     - these resources must be explicitly locked/unlocked
+ *                     - locked resources prevent (obviously) the kernel
+ *                       context from being unloaded.
+ *                     - drivers using these resource directly issue their own
+ *                       GRU instruction and must wait/check completion.
+ *
+ *               When these resources are reserved, the caller can optionally
+ *               associate a wait_queue with the resources and use asynchronous
+ *               GRU instructions. When an async GRU instruction completes, the
+ *               driver will do a wakeup on the event.
+ *
  */
+
+
+#define ASYNC_HAN_TO_BID(h)    ((h) - 1)
+#define ASYNC_BID_TO_HAN(b)    ((b) + 1)
+#define ASYNC_HAN_TO_BS(h)     gru_base[ASYNC_HAN_TO_BID(h)]
+
 #define GRU_NUM_KERNEL_CBR     1
 #define GRU_NUM_KERNEL_DSR_BYTES 256
 #define GRU_NUM_KERNEL_DSR_CL  (GRU_NUM_KERNEL_DSR_BYTES /             \
@@ -99,20 +145,6 @@ struct message_header {
 #define HSTATUS(mq, h) ((mq) + offsetof(struct message_queue, hstatus[h]))
 
 /*
- * Allocate a kernel context (GTS) for the specified blade.
- *     - protected by writelock on bs_kgts_sema.
- */
-static void gru_alloc_kernel_context(struct gru_blade_state *bs, int blade_id)
-{
-       int cbr_au_count, dsr_au_count, ncpus;
-
-       ncpus = uv_blade_nr_possible_cpus(blade_id);
-       cbr_au_count = GRU_CB_COUNT_TO_AU(GRU_NUM_KERNEL_CBR * ncpus);
-       dsr_au_count = GRU_DS_BYTES_TO_AU(GRU_NUM_KERNEL_DSR_BYTES * ncpus);
-       bs->bs_kgts = gru_alloc_gts(NULL, cbr_au_count, dsr_au_count, 0, 0);
-}
-
-/*
  * Reload the blade's kernel context into a GRU chiplet. Called holding
  * the bs_kgts_sema for READ. Will steal user contexts if necessary.
  */
@@ -121,17 +153,23 @@ static void gru_load_kernel_context(struct gru_blade_state *bs, int blade_id)
        struct gru_state *gru;
        struct gru_thread_state *kgts;
        void *vaddr;
-       int ctxnum;
+       int ctxnum, ncpus;
 
        up_read(&bs->bs_kgts_sema);
        down_write(&bs->bs_kgts_sema);
 
        if (!bs->bs_kgts)
-               gru_alloc_kernel_context(bs, blade_id);
+               bs->bs_kgts = gru_alloc_gts(NULL, 0, 0, 0, 0);
        kgts = bs->bs_kgts;
 
        if (!kgts->ts_gru) {
                STAT(load_kernel_context);
+               ncpus = uv_blade_nr_possible_cpus(blade_id);
+               kgts->ts_cbr_au_count = GRU_CB_COUNT_TO_AU(
+                       GRU_NUM_KERNEL_CBR * ncpus + bs->bs_async_cbrs);
+               kgts->ts_dsr_au_count = GRU_DS_BYTES_TO_AU(
+                       GRU_NUM_KERNEL_DSR_BYTES * ncpus +
+                               bs->bs_async_dsr_bytes);
                while (!gru_assign_gru_context(kgts, blade_id)) {
                        msleep(1);
                        gru_steal_context(kgts, blade_id);
@@ -203,6 +241,114 @@ static void gru_free_cpu_resources(void *cb, void *dsr)
        preempt_enable();
 }
 
+/*
+ * Reserve GRU resources to be used asynchronously.
+ *   Note: currently supports only 1 reservation per blade.
+ *
+ *     input:
+ *             blade_id  - blade on which resources should be reserved
+ *             cbrs      - number of CBRs
+ *             dsr_bytes - number of DSR bytes needed
+ *     output:
+ *             handle to identify resource
+ *             (0 = async resources already reserved)
+ */
+unsigned long gru_reserve_async_resources(int blade_id, int cbrs, int dsr_bytes,
+                       struct completion *cmp)
+{
+       struct gru_blade_state *bs;
+       struct gru_thread_state *kgts;
+       int ret = 0;
+
+       bs = gru_base[blade_id];
+
+       down_write(&bs->bs_kgts_sema);
+
+       /* Verify no resources already reserved */
+       if (bs->bs_async_dsr_bytes + bs->bs_async_cbrs)
+               goto done;
+       bs->bs_async_dsr_bytes = dsr_bytes;
+       bs->bs_async_cbrs = cbrs;
+       bs->bs_async_wq = cmp;
+       kgts = bs->bs_kgts;
+
+       /* Resources changed. Unload context if already loaded */
+       if (kgts && kgts->ts_gru)
+               gru_unload_context(kgts, 0);
+       ret = ASYNC_BID_TO_HAN(blade_id);
+
+done:
+       up_write(&bs->bs_kgts_sema);
+       return ret;
+}
+
+/*
+ * Release async resources previously reserved.
+ *
+ *     input:
+ *             han - handle to identify resources
+ */
+void gru_release_async_resources(unsigned long han)
+{
+       struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
+
+       down_write(&bs->bs_kgts_sema);
+       bs->bs_async_dsr_bytes = 0;
+       bs->bs_async_cbrs = 0;
+       bs->bs_async_wq = NULL;
+       up_write(&bs->bs_kgts_sema);
+}
+
+/*
+ * Wait for async GRU instructions to complete.
+ *
+ *     input:
+ *             han - handle to identify resources
+ */
+void gru_wait_async_cbr(unsigned long han)
+{
+       struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
+
+       wait_for_completion(bs->bs_async_wq);
+       mb();
+}
+
+/*
+ * Lock previous reserved async GRU resources
+ *
+ *     input:
+ *             han - handle to identify resources
+ *     output:
+ *             cb  - pointer to first CBR
+ *             dsr - pointer to first DSR
+ */
+void gru_lock_async_resource(unsigned long han,  void **cb, void **dsr)
+{
+       struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
+       int blade_id = ASYNC_HAN_TO_BID(han);
+       int ncpus;
+
+       gru_lock_kernel_context(blade_id);
+       ncpus = uv_blade_nr_possible_cpus(blade_id);
+       if (cb)
+               *cb = bs->kernel_cb + ncpus * GRU_HANDLE_STRIDE;
+       if (dsr)
+               *dsr = bs->kernel_dsr + ncpus * GRU_NUM_KERNEL_DSR_BYTES;
+}
+
+/*
+ * Unlock previous reserved async GRU resources
+ *
+ *     input:
+ *             han - handle to identify resources
+ */
+void gru_unlock_async_resource(unsigned long han)
+{
+       int blade_id = ASYNC_HAN_TO_BID(han);
+
+       gru_unlock_kernel_context(blade_id);
+}
+
 /*----------------------------------------------------------------------*/
 int gru_get_cb_exception_detail(void *cb,
                struct control_block_extended_exc_detail *excdet)
index 747ed31..d60d34b 100644 (file)
@@ -146,4 +146,55 @@ extern void *gru_get_next_message(struct gru_message_queue_desc *mqd);
 extern int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa,
                                                        unsigned int bytes);
 
+/*
+ * Reserve GRU resources to be used asynchronously.
+ *
+ *     input:
+ *             blade_id  - blade on which resources should be reserved
+ *             cbrs      - number of CBRs
+ *             dsr_bytes - number of DSR bytes needed
+ *             cmp       - completion structure for waiting for
+ *                         async completions
+ *     output:
+ *             handle to identify resource
+ *             (0 = no resources)
+ */
+extern unsigned long gru_reserve_async_resources(int blade_id, int cbrs, int dsr_bytes,
+                               struct completion *cmp);
+
+/*
+ * Release async resources previously reserved.
+ *
+ *     input:
+ *             han - handle to identify resources
+ */
+extern void gru_release_async_resources(unsigned long han);
+
+/*
+ * Wait for async GRU instructions to complete.
+ *
+ *     input:
+ *             han - handle to identify resources
+ */
+extern void gru_wait_async_cbr(unsigned long han);
+
+/*
+ * Lock previous reserved async GRU resources
+ *
+ *     input:
+ *             han - handle to identify resources
+ *     output:
+ *             cb  - pointer to first CBR
+ *             dsr - pointer to first DSR
+ */
+extern void gru_lock_async_resource(unsigned long han,  void **cb, void **dsr);
+
+/*
+ * Unlock previous reserved async GRU resources
+ *
+ *     input:
+ *             han - handle to identify resources
+ */
+extern void gru_unlock_async_resource(unsigned long han);
+
 #endif                 /* __GRU_KSERVICES_H_ */